Fix Phase 2 Task Validation: iterate all 3 tasks in inference.py

Update CLAUDE.md to reflect pre-submission upgrade sprint state

Clean pyproject.toml package configuration and remove fastmcp dead dep

Trim task_definitions.py to match minimal passing pattern

Switch server/app.py to create_fastapi_app and simplify /tasks response

Simplify openenv.yaml to match passing submission schema

Fix Phase 2 'Not enough tasks with graders' — add canonical TASKS registry

Add Claude context files: CLAUDE.md + skill + slash commands

Fix Phase 2: add GET /tasks and POST /grader endpoints

Add uv.lock for reproducible dependency resolution

Re-enable Gradio web UI at / for friendly browser-facing Space page

Update README for canonical OpenEnv structure and submission spec

Update tests to use inline heuristic helper after inference.py rewrite

Add scripts/validate-submission.sh — runs the 3 grader checks locally

Rewrite inference.py to match hackathon submission spec exactly

Rewrite client.py as EnvClient subclass with WebSocket protocol

Replace MCPEnvironment with canonical Environment base class

Refactor models.py: inherit from openenv Action/Observation/State

Add EXPOSE 8000 so HF Spaces port detection works

Set app_port=8000 so HF Spaces routes to FastAPI

Add README with environment description, action/observation spaces, and baseline scores

Add inference.py with heuristic and LLM agent modes, calibration tests, and baseline results

Add easy/medium/hard task scenarios and programmatic grader

Add MCPEnvironment subclass with 5 dispatcher tools and FastAPI server

Add DispatchSimulation engine, geometry helpers, caller text templates, and observation renderer

Add Pydantic models and clinical survival curves (cardiac, trauma, stroke, fire)

Initial scaffold: pyproject, Dockerfile, openenv.yaml